AJava-basedintelligentadvisorforselectingacontext- - …kclarke/Geography232/cooper_jar… · · 2005-02-02Computers&Geosciences30(2004)1003–1018 AJava-basedintelligentadvisorforselectingacontext-appropriatespatialinterpolationalgorithm$,$$

ARTICLE IN PRESS

0098-3004/$ - se

doi:10.1016/j.ca

$Paper prese

University of S$$Editorial�CorrespondE-mail addr

Computers & Geosciences 30 (2004) 1003–1018

www.elsevier.com/locate/cageo

A Java-based intelligent advisor for selecting a context-appropriate spatial interpolation algorithm$,$$

William Coopera,�, Claire Jarvisb

aSAIC Ltd, Campus 1, Aberdeen Science and Technology Park, Bridge of Don, Aberdeen, UKbDepartment of Geography, University of Leicester, University Road, Leicester, UK

Accepted 18 July 2004

Abstract

The choice of an appropriate interpolation technique is an imposing task. There are many techniques to choose from

depending on the characteristics of the phenomenon, the purpose of the study and also the expertise of the user. This

paper describes the implementation and architecture of a Java-based intelligent advisor to assist a generic user, ranging

from the casual to the specialist, in selecting the interpolator most appropriate for a given task and data set. The

software integrates procedural knowledge from disparate sources through an expert system shell, and uses multiple

sensory channels to present and abstract contextual knowledge from the user regarding both the data and the task at

hand. The system has been tested for use in a meteorological domain in the first instance. The interpolation methods

currently assessed by the system are multiple forms of kriging, thin plate smoothing splines, inverse-distance weighting

and trend surface/polynomial analysis.

r 2004 Elsevier Ltd. All rights reserved.

Keywords: Interpolation selection; Expert systems; Java and Jess; AI; Meteorology

1. Introduction

Spatial interpolation methods are widely used in

creating continuous environmental data sets from a

network of sparsely sampled points. In particular, they

have been employed to build continuous representations

of terrain, soil composition, terrestrial and atmospheric

pollution and climate variables (Heuvelink and Webster,

2001; Hutchinson and Gallant, 1999; Jarvis and Stuart,

2001a; Oliver and Khayrat, 2000). The task of inter-

polator selection, of choosing a credible spatial model to

e front matter r 2004 Elsevier Ltd. All rights reserve

geo.2004.07.007

nted at the GIS Research UK conference,

heffield, April 2002.

handling by S.M.Wise.

ing author. Tel.: +441224 333 600.

ess: [email protected] (W. Cooper).

suit the data, is however non-trivial. The appropriate-

ness of a particular method from the wide range

available is dependent on a number of factors, such as

the characteristics of the phenomenon and the purpose

of the study. Judgement is required of the user that

requires familiarity both with the spatial associations in

the domain data and the functions and assumptions

behind the methods of interpolation. Thus the casual

user who has received no formal instruction in

GIScience or spatial statistics may find it difficult to

judge the applicability of a method, whilst the data

analyst who casually dismisses the assumptions in

interpolation methods is the one most likely to ‘‘misuse

them or to be misled by them’’ (Kitanidis, 1999, p. 7–8).

The challenge in the selection of an interpolator is to

identify an adequate method with appropriate para-

meters that best suits the user’s expertise, data and task

d.

www.elsevier.com/locate/cageo

ARTICLE IN PRESS

SpatialAnalysis

ExpertKnowledge

Rules

W. Cooper, C. Jarvis / Computers & Geosciences 30 (2004) 1003–10181004

(Mitas and Mitasova, 1999). This paper addresses the

implementation of an ‘intelligent’ advisor to assist users

of GIS with the problem of selecting the most

‘appropriate’ interpolation method on a task and data-

dependent basis.

InformationVisualization

Knowingyour data

Cognition

"Intelligent"ExpertSystem

Fig. 1. Filling gap: major methodological elements of software

required to provide user with ‘intelligent’ support.

2. Background

Considerable polarization exists within the current

literature for selecting an appropriate interpolation

method, between extended exploratory data analysis

(EEDA) on one side (Bucher, 1998) and simple knowl-

edge-based methods on the other (Dimitrakopoulos,

1993; Maslyn, 1987). Both of these approaches indivi-

dually address only a limited, if valuable, part of the

overall problem and so limit the practical support they

can provide for the user. A gap exists in supporting the

selection of an interpolation method when evaluating

the area of model selection and procedures as a whole. In

other words, within GIScience to date there is no one

tool that encapsulates the experience of an expert tutor

and supports decisions throughout all key tasks when

selecting which interpolation method to use under a

certain set of circumstances.

In attempting to fill this gap, we firstly consider the

potential synergy among various automated and visual

statistical modules that might be required. Just as the

use of linked windows has assisted the discerning of

relationships between GI data using visual expression

(Bastin et al., 2002; Dykes, 1997), we argue that similar

links are required to bring together the distinct statistical

analyses needed within the interpolation selection

process on an automated basis. Furthermore, to support

the user in selecting an interpolator, and following Mitas

and Mitasova (1999), we argue that intelligence should

be incorporated into the selection procedure. In other

words, we consider the software for the task of selecting

an interpolation technique discussed in this paper to be a

symbiotic construction of spatial analysis (exploratory

data analysis/statistics), exploratory information visua-

lization techniques and expert knowledge (Fig. 1).

As we shall demonstrate, the software seeks to

balance deductive and inductive methods, in this case

using Expert System (ES) technology in addition to

assisted data analysis, to represent a large body of expert

knowledge regarding the functions and assumptions of

interpolation methods. The flexibility to use both

qualitative symbolic and quantitative numerical meth-

ods as required in one application provides great

potential to deliver an understanding of the initial

assumptions and methods used in the spatial interpola-

tion process. It further permits the system to be

responsive to the information, task and context at hand.

Recent developments in intelligent systems emphasise

the importance of context when developing flexible ES

in order to define when knowledge should be considered

for practical decision support (Brezillon, 1999). We

advocate that such context might be achieved here

through an approach that accompanies knowledge of

the GIScience procedures themselves with both the

users’ knowledge of their domain and a visual and

numerical analysis of the data. In a previous paper, we

outlined the theoretical constructs of artificially intelli-

gent support and how knowledge from interpolation

experts could be conceptually structured (Jarvis et al.,

2003). This paper serves to discuss the implementation

of these concepts into a software system and to provide

a walkthrough of the system in operation. Here we also

discuss the verification and validation of our strategy for

assisting the users of GIS in their selection of interpola-

tion methods that are appropriate to the task and data.

3. Scope of the software

GIS and statistical software contain diverse options

for spatial analysis, but a user of one system in

particular is unlikely to have a complete range of

possible tools at hand. Thus, for the broadest adoption

of first initial generation intelligent GI support modules,

we advocate a loosely coupled approach (Fig. 2) that is

able to serve the many GIS and statistical packages

currently commercially available, each with their differ-

ent scripting and interfacing capabilities. That is, input

point sample data may be imported to the advisor

software via simple ASCII files and simple parameter

files may be exported.

We also designed the intelligent interpolation advisor

to choose between interpolation methods that are

ARTICLE IN PRESS

Informatics ModuleDecision SupportModule

InformationVisualization

ExploratoryStatistics

InferenceEngine

UserCognition

Spatial AnalysisModule

Knowledge &Parameter support

Expert SystemModule

InformationEnrichment

User Interface

ExternalInterpolation

ProgramData Set

1.6 32.6 0.6 0.14.1 44.5 0.4 -990.6 64 0.5 0.12.4 71.8 0.3 0.37.8 3.5 0.4 0.1

ExpertRules

Expert System Interpolant Selector

Facts

Fig. 2. Interpolation, spatial data and an ‘‘intelligent’’ expert system: their integration.

3.4 STEP-WISECONSTRUCTION

3.1 NEEDS ASSESSMENT

Software Support

Implement

Extend

Verification

Yes

No

The acquisition processPhase

Teaching

Design Scenario

Encapsulating ModalityJava Classes and

Programming

3.5 VALIDATION

3.2 INTEGRATING THEEXPERT SYSTEM

3.3 KNOWLEDGEACQUISITION

W. Cooper, C. Jarvis / Computers & Geosciences 30 (2004) 1003–1018 1005

commonly available in many packages and that are

widely used in many domains. These include, in the first

instance, trend surface analysis, inverse distance weight-

ing (IDW), thin plate splines and the kriging family. The

system is designed to suggest a primary and secondary

interpolation method suitable for the task at hand.

Where a method is suggested to which a user has no

access, the intention is that firstly the suggested

alternative method might be used and secondly, where

this performs poorly, the user might be prompted to

access additional interpolation software. The software

itself does not provide interpolation functions, nor

pander to the functionality of one proprietary GIS or

one theoretical camp. Its rationale is rather to advise on

which methods a user might best consider based on the

collected evidence, expecting that a user will have

expertise in their particular domain to judge the realism

of the surface(s) suggested by the selection software but

constructed within a proprietary GIS and also the ability

to carry out independent sampling of the surface(s) for

accuracy.

3.6 EVALUATIONSTRATEGY

ValidationDomainSphere

Fig. 3. Main phases in software development.

4. Implementation: software design and system

architecture

For reference, Fig. 3 outlines the main phases in the

development of the system as a whole. The software

design scenario process established that the performance

and functionality needs of a modular system would best

be managed using Javar. Java from SunTM (Shannon

et al., 2000) is a platform independent object orientated

language which is suitable for calling external programs

(interpolation algorithms) for providing openess as a

portal for future implementation of the module over the

ARTICLE IN PRESSW. Cooper, C. Jarvis / Computers & Geosciences 30 (2004) 1003–10181006

World Wide Web and for its interconnectivity with ES

shells. In addition, the growing availability of modular,

re-useable classes for both spatial analysis tasks

(Faulkner, 2002) and interface development more

generally made Java an attractive choice.

System architecture was conceived with three main

underlying components, brought together within a

multi-modal (Egenhofer, 1996) user interface (Section

4.1). These include an expert systems module (Section

4.2) an informatics module (Section 4.3) and a decision

support module (Section 4.4).

4.1. User interface

The main challenge in designing an interface is to

‘‘assist in the visualization and manipulation of complex

domains’’ (Lypher and Stelzner, 1991, p. 404), in this

case the theoretical world of interpolation as expressed

within an expert system. In drawing together the textual,

visual and graphical channels that are required for the

desired partnership between spatial analysis (explora-

tory data analysis/statistics), expert knowledge and

exploratory information visualization, we implicitly

adopt the advantages of multi-modal channels (Egen-

hofer, 1996) to support the user in making decisions

based on various criteria. That is, an interface that

appeals to multiple elements of their perception in order

to explain a topic, for example by providing written text

in the form of a question and an accompanying picture

explaining the concept in an alternative fashion.

Combining these complementary channels should ex-

pose to the user various aspects of the data and

interpolation procedures.

In the field of human computer interfaces, it has been

suggested that users are the most variable component

(Shneiderman, 1998). This arises from the variety of user

preferences, varying familiarity with a system or concept

(e.g. interpolation), previous experience and expecta-

tions, and an understanding and knowledge of terms

used. In order to minimize these variables, we therefore

created a ‘‘look’’ for the interpolation selection assistant

based upon the familiar ‘‘Desktop’’ metaphor.

The main work area of the prototype interface is

therefore analogous to the ‘‘Desktop’’ (Fig. 4). This

Desktop allows for informatics, statistics, X–Y plot,

variogram, histogram (not shown) and distribution

curve internal frames to be generated. The Informatics

frame is automatically made once the program is run. It

is subdivided into a left- and right-hand side, with two

buttons on the bottom. The left-hand side presents to

the user questions, which are answered by pressing either

the yes or no buttons provided. On the right-hand side of

the frame, information visualization is presented, to aid

the user in understanding questions, which may be

highly contextual. The parameter module is found in the

‘‘Menubar’’ at the top of the desktop. The Menubar

provides further assistance, allowing other frames to be

generated when required, as shown in Fig. 5.

The Menubar consists of seven main items: Load, File,

Exploratory Analysis Module, Parameter Analysis Mod-

ule and the Info drop down box. Working through these

in order, Load allows a variety of data input file formats

to be loaded, so long as they are comma delimited and in

an X, Y, dependent variable, covariate format where the

covariates are optional. The standard loading facilities

of the operating platform are adopted by Java. In the

above example, Java is being run on the Windowss 95

platform and the associated load box is presented. File

controls the Informatics frame. It provides a drop down

menu that can clear out the frame contents, and allows

the user to exit the inference process. A restart button

will allow the inference process to be repeated if the user

wishes.

The task of interpolation selection requires the user to

use judgement and make decisions, for which additional

familiarity with the data and interpolation methods over

and above that provided automatically by the software

is helpful. Thus, the Exploratory Analysis Module drop

down box handles various descriptive statistics and

correlation calculations, which are presented to the user

in the Statistics Frame. Histograms, X–Y plots and

distribution curves are calculated in their associated

frames. Finally, the Parameter Analysis Module is used

after the Informatics frame gives a recommendation.

This allows the user to determine the parameters

required to carry out an interpolation method, for

example the most appropriate model with which to fit a

variogram.

4.2. Expert systems module

Generalizations in regard to expert systems as a whole

have been discussed at length by Duda and Shortliffe

(1983) and Hayes-Roth et al. (1983) among others, and

are summarized in relation to this software in Table 1.

These include for example the modularity of rules, useful

for different geographical or application domains, and

the separation of control from knowledge such that

alternative and more flexible, softer forms of control can

be explored in future with relative ease. Thus, the

decision to implement this prototype software around

expert system architecture has many advantages in

regard to the future stability and expansion of the

system. This is regardless of the simplicity of the control

strategy by which we structure knowledge in this

software prototype.

4.2.1. Control strategy

A control strategy is vital to pull together all the

components of the ES. There are many types of control

strategies, including frame based, induction, goal agenda

and hybrid approaches. It is, however known that where

ARTICLE IN PRESS

Informatics Frame Menubar X-Y Plot Frame

Statistics Frame Variogram Frame Distibution Curve Frame

Fig. 4. System interface ‘‘desktop’’.


a global problem solving approach is used by the expert,

a strategy using rules best matches the problem (Durkin,

1994). Following this advice, we concluded that in the

case of interpolation a rule-based control strategy would

offer a good means of structuring the knowledge of

experts. In this case, expert knowledge for interpolation

is held as ‘rules of thumb’ which specify a set of actions

to be performed for a given situation in the form ‘if

/conditionS then /conclusionS’. A rule-based ap-

proach allows the acquired knowledge to be external to

the program, in this system held in a text file, so the

expert knowledge can be easily expanded, archived or

viewed.

4.2.2. Organizing the rules: conceptual and formal

structuring of knowledge

Much of the work in developing the rule-based

approach is focused on accumulating all of the knowl-

edge necessary to resolve a problem. The method used to

elicit and acquire knowledge of the interpolation

selection process involved close collaboration with the

domain experts and extensive research of the literature

as discussed more extensively elsewhere (Jarvis et al.,

2003). Out of this previous work came a common

sequence in the way in which an expert often asks

questions or employs knowledge at each stage of the

decision making process in selecting an interpolation

method. This provides an important context for the

process of implementation and so is expanded upon

here.

In the first instance, the interpolation selection process

needs to review a study’s objectives and the user’s

accuracy requirements. We note for example the

important divide within project objectives between

whether a data surface is used for analysis or is required

for aesthetic purposes. As Declercq (1996, p. 128) notes,

‘‘highly accurate interpolations do not always produce

realistic spatial patterns’’. Moreover, notable problems

in the representation of visual pattern have been noted

with particular interpolators such as IDW where data

ARTICLE IN PRESS

Fig. 5. Menu bar functionality.

Table 1

Advantages of use of an expert system approach in this decision-tree-based version of intelligent interpolation advisor (After Duda and

Shortliffe, 1983; Hayes-Roth et al., 1983)

Advantages Explanation

Modularity of rules Modularity makes it possible to structure the knowledge of the domain being studied as a series

of distinct production rules collected together for application within a particular domain under

consideration. In this context, this might apply to differing geographical or application domains.

Separation of control form

knowledge

The separation of control from the knowledge is a valuable feature, which permits expansion of

the knowledge base without changing the systems inference engine. Here, future versions of this

prototype might include softer, fuzzy rules for example, managed using backward chaining.

Consistency checking and

verification

The rigid structure of the rules allows for easier checking of any conflicts within the rules. This

advantage is an important improvement over a hand-coded decision tree implementation.

Growth of intelligence As more rules can be added easily to the knowledge base, the rule-based system promotes

development and thus the systems level of intelligence increases to expand to a wider problem

domain. This is valuable, given the wide range of interpolation methods used across multiple

domains.


are sparse and trend has not been accounted for in the

data prior to interpolation (Collins and Bolstad, 1996;

Declercq, 1996). At the root of the diagnostic process

therefore the purpose or context is established from the

user of the software, which directly influences the

direction of the subsequent problem solving process.

ARTICLE IN PRESSW. Cooper, C. Jarvis / Computers & Geosciences 30 (2004) 1003–1018 1009

This division, and others similar further down the chain

of reasoning, led us to formalize expert knowledge

contained within the software as a binary decision tree

(Fig. 6). All nodes within the tree contain a question to

which the user is asked for a [yes] or [no] outcome.

Further down the knowledge structure, we identify

questions that draw on the domain expert’s under-

standing of the phenomenon being interpolated and the

existence and nature of any digital data that they might

have to hand. Would a ‘smooth’ or ‘bumpy’ outcome

more suit their expectations, for example? The meteor-

ological user would know that monthly average or

climate normal surfaces will be smoother than surfaces

of daily meteorological variables, and that a precipita-

tion surface for Britain is likely to be more spatially

fragmented than a corresponding temperature surface.

Further questions in this vein assess whether associated

data are available at co-located points to the dependent

variable or as continuous a surface such as elevation, if

indeed at all. These ‘top down’ concepts commonly form

the earliest questions that an expert might ask in

diagnosing which interpolator to use, and form part of

a previously implicit knowledge structure.

From this point in the decision tree structure, more

technical questions need to be raised and addressed. For

example, how large is the sample?, are the data normal?,

if anisotropy is expected, what is the ratio and direction

of greatest variation indicated by the data?. Moving

through this range of considerations relies on expert

knowledge of the procedures involved in selecting an

interpolation method, and invoking this procedural

knowledge is reliant on local context. An important

function of the intelligent software is to ask questions

and provoke user interaction to answer as many of these

technical questions as possible ‘behind the scenes’.

Classifying an appropriate interpolation method using

the decision tree might thus occur as follows for the

example outline in Fig. 6. Firstly, beginning with the

root node (1), we determine if the new task is either a

visualization [yes] or analytical problem [no]. In our

1

2

45 6

3

7

8 9 10 11 2 13

ROOT NODE

BRANCH

NODE

LEAF

Traverse Direction

Example

14 15

16 17 18 19

Fig. 6. Binary decision tree.

study case the outcome is [yes] so the left branch is

traversed. The next node (2) is a non-leaf node, so a new

question is tested, i.e. ‘‘is the data strongly related to

another variable?’’ The answer to this is [yes] and the

corresponding branch is traversed. The next node (4) is

also a non-leaf node so a further question is asked, ‘‘Do

you have digital data for this variable?’’. The answer to

this is [no] which leads to node (9) where the question

‘‘are you expecting a smooth [yes] or bumpy surface

[no]’’. The answer to this is [yes] and the corresponding

branch is traversed to a leaf where a classification of

Regularized Spline Tension or Smoothing Splines is

given. A decision tree traverse such as above might

provide the impression that only a handful of rules have

been encoded within the expert system. It is important to

note that, while a handful of rules are directed at the

user at any one time the overall number of questions

that could be asked according to the problem at hand is

considerably higher; the system encapsulates almost 50

rules at the present time with scope for further

expansion. It is also worth reiterating at this juncture

that the user can employ statistical and visual tools from

the overall system toolbox to support the reasoning

process.

4.2.3. Encoding: physical structuring of knowledge

We engaged Jessr as a means to encode the

requirements of the control strategy and procedural

knowledge. Jess from Friedman-Hill (1997) is a rule-

based expert system shell and scripting language written

entirely with the Java language. It is based on the kernel

of the CLIPS-ES (Giarratano and Riley, 1989), which

has broad acceptance both in the AI literature and more

specifically within GIScience. It utilizes the Rete

algorithm (Forgy, 1982) inference engine.

The expert knowledge conceptually formulated in the

binary decision tree and in part described in Table 2 was

physically implemented into a tree-structured node

network of Jess IF–THEN statements called production

rules, illustrated within Table 3. While these structures

combined could have been implemented in a procedural

language, we used the Jess formalization methodology

for its ability to separate knowledge from rules to aid

extensibility, useful for scaling the prototype to addi-

tional domains and methods in the future. The produc-

tion rules have a left-hand side (LHS) and a right-hand

side separated by the ‘‘) ’’ symbol, which can be read

as THEN. During the system operation an initial root

question is asked and stored in the LHS of the

production rule. In answering by console interaction,

the user declares knowledge and these ‘new facts’ are

placed in working memory. Rete processes the LHS

knowledge and the RHS action for conditions that can

be satisfied by working memory. When a rule’s LHS

matches a fact, the RHS is then activated, the current

node is retracted and Rete fires to the next node in the

ARTICLE IN PRESS

Table 2

Network of nodes: decision tree encoding

(node (name root) (type decision) (question ‘‘Do you wish to generate a surface for

Visualization (i.e. just a quick aesthetic surface) [PRESS YES] or a method

that produces values for the purpose of analysis [PRESS NO] ? 1’’) (yes-node

node2) (no-node node3) (answer nil))

(node (name node2) (type decision) (question ‘‘Is the surface to be generated

strongly related to another variable? 2’’) (yes-node node4) (no-node node5)

(answer nil))

(node (name node4) (type decision) (question ‘‘Do you have digital data for this

variable? 3’’) (yes-node node8) (no-node node9) (answer nil))

(node (name node9) (type decision) (question ‘‘Are you expecting the generated

surface to be Smooth [PRESS YES] or Bumpy [PRESS NO]? 7’’) (yes-node node

18) (no-node node 19) (answer nil))

(node (name node 18) (type answer) (question nil) (yes-node nil) (no-node nil)

(answer

BEST_METHOD_#1=reguralised_spline_tension_OR_#2=smoothing_splines_0))


network. This determines which questions are asked and

in what order, and eventually leads to a recommenda-

tion. For example, within the node network and starting

with the root node, the system asks a decision node a

question and reads the answer (only ‘‘yes’’ or ‘‘no’’). The

current node is then retracted, and the reasoning

proceeds downwards to the next node using either the

yes or no branch, as relevant, and onwards through the

decision tree structure. Fig. 7 provides examples of node

activation in the interpolation context at the upper levels

of the node network tree involved with the user’s

knowledge of the task and supporting data.

4.3. Informatics module

The informatics module promotes semantic under-

standing, contextual recognition and statistical feature

abstraction, with the aim of helping users to understand

the potentially complex questions asked of them during

the course of a ‘conversation’ with the software. These

questions prompt the user to carry out the inference

process and procedure used by an expert. However,

since a domain expert retains a large number of image

symbolic abstractions relating to their expert knowledge,

graphical representations are presented which relate to

individual questions. The objective of this is to inspire

and connect with the thought processes of the user more

efficiently than using communication by text alone. For

example, the question ‘do you expect the surface variable

to be smooth or bumpy?’ is accompanied by pictures of

two surfaces exhibiting very different levels of auto-

correlation. The effect of answering a question relating

to the desirability or otherwise, from a sample accuracy

perspective, of retaining exact data measurements within

the surface is illustrating by a curve showing a smooth

surface but with overshooting or undershooting sample

points. The expectation of anisotropy within the data is

similarly best explained to the geostatistical layperson in

diagrammatic terms or the danger is that the user of the

software is simply over-awed by the questions asked,

rather than supported as is the intention.

To further present contextual clues we encourage the

user to explore their data using preliminary spatial and

statistical investigations. Exploring data visually by

using scattergrams and distribution graphs such as

histograms and variograms is acknowledged to be

beneficial to cognition, and allows the user to be

informed of their data more effectively (Fotheringham

et al., 2000; Kitanidis, 1999). An example of this facility

in action would be the ability to check for the

significance of a linear regression between dependent

and independent variables if the domain user was not

certain of this a priori.

4.4. Decision support module

Having been supported in the selection of a range of

potentially satisfactory interpolation methods, the user

still requires assistance with the setting of parameters to

control the generation of surfaces. The majority of GIS

systems leave the selection of parameters entirely in the

hands of the user, and an ill-judged choice may give rise

to a surface that is not intended (Bonham-Carter, 1994).

The rationale behind a parameter module is to aid the

user by suggesting such parameters as the variogram

model, range and sill in the case of ordinary kriging or

an appropriate distance decay rate for IDW (Table 4).

Inferred solutions will not necessarily be unique

within the context of interpolation. Bounding an ill-

defined problem such as selecting an interpolation

method is an approach that we adopt from spatial

decision support systems (Cameron and Abel, 1996).

ARTICLE IN PRESS

Table 3

Production system rules

(deftemplate node (slot name) (slot type) (slot question) (slot yes-node) (slot no-node) (slot

answer))

(defrule initialize-1 (not (node (name root)))

=4(load-facts ‘‘knowledge.dat’’) (assert (current-node root)))

(defrule initialize-2

(declare (salience 100))? fact o- (next-gensym-idx ?idx)

=4(retract ?fact)(setgen ?idx))

(defrule ask-decision-node-question

?node o- (current-node ?name) (node (name ?name) (type decision) (question?question)) (not (answer

?))

=4(printout t ?question crlf) (assert (answer (read))))

(defrule bad-answer?answer o- (answer~yes&~no)

=4(retract ?answer))

(defrule proceed-to-yes-branch

?node o- (current-node ?name) (node (name ?name) (type decision)

(yes-node ?yes-branch)) ?answer o- (answer yes)

=4(retract ?node ?answer) (assert (current-node ?yes-branch)))

(defrule proceed-to-no-branch

?node o- (current-node ?name) (node (name ?name) (type decision)

(no-node ?no-branch)) ?answer o- (answer no)

=4(retract ?node ?answer) (assert (current-node ?no-branch)))

(defrule print answer

?node o- (current-node ?name) (node (name ?name) (type answer) (answer ?value)) (not (answer ?))

=4(printout t‘‘- - - - - - - - - - - - - - - - - - -end of inference- - - - - - - - - - -

(reset)

(run)

Task isvisualisation

18

2Data strongly

related to anothervariable

4Dont have datafor variable

Expecting asmooth surface

9

Classified asRegularised

SplineTension

Fired on NODE 2

Fired on NODE 4

Fired on NODE 9

Activated

Fig. 7. Sample rule activation.



The intention here is that the system will present to the

user the most and least appropriate methods. This

flexibility is also important in the context of the loosely

coupled approach, where only some of the feasible

options may be available to the user within the context

of their particular GIS.

4.5. Encapsulating modality: a description of the Java

classes

The above modules are tightly encapsulated in object-

orientated relationships using the following Java classes:

re-usable classes in JDK 2, the ES Jess package and a

newly composed interpolation-advisor package. The

import statement was used at the top of new code to

invoke the prewritten and precompiled classes. New

classes were constructed through inheritance and com-

position and were as atomic as possible, sub dividing the

problem between subsystem functionality and system

modality. In part the main components were declared

into the ‘‘StatisticsModule’’ class for exploratory data

analysis i.e. to communicate the data set characteristics,

the ‘‘ShowImage’’ class to aid question cognition, the

‘‘Data SetLoader’’ class to load the data, and the

‘‘ConsolePanel’’ class to build the Graphical User

Interface, which presents the information to the user.

To complete the task of class design we specified ‘‘use-

cases’’, which described the internal and external

behaviour of the system allowing internal objects to be

determined (Erdmann and Studer, 1998), for example:

the ‘‘StatisticsModule’’ class needed to interface with the

‘‘ConsolePanel’’ class in order to present the user with

statistics. Such communication was achieved by defining

an object of the ‘‘StatisticsModule’’ in the ‘‘ConsolePa-

nel’’ through constructors and suitable arguments.

The class Jess Console was the main() method i.e. the

entry point to the software. This was declared public and

static to negate the necessity of constructing an instance

of the class and thus allow execution. The system class

architecture is shown in Fig. 8 and detailed java class

documentation (JavaDoc) in Cooper (2001).

Table 4

Parameters required of user for interpolation tasks

Method Associated parameters required by

GIS

General Trend element (e.g. regression

parameter, order of trend)

Geo-statistics Variogram model type, variogram

parameter, anisotropy

I.D.W No of neighborhood points, power

parameter

Splines Tension and smoothing parameters

5. Scenario walkthrough: selecting a suitable

interpolation for precipitation

The interplay between these modules can be seen more

directly by a walkthrough of the software. In this case,

an M.Sc. student using interpolation methods for the

estimation of precipitation was asked to perform a

scenario walkthrough using daily precipitation data for

65 meteorological stations in Peru (i.e. a data set not

part of the validation). A meteorological domain was

chosen for this walkthrough, and for the later validation

and usability tests (Section 6) since spatially distributed

estimates of meteorological data are becoming increas-

ingly important as inputs to environmental models.

Additionally, there exists a large body of published work

and local knowledge and data (e.g. Goovaerts, 1999;

Hutchinson, 1995; Jarvis and Stuart, 2001b) which could

be used both to develop appropriate rules and to test

empirically the relative suitability of the different

interpolators in this domain. The interpolation methods

most commonly used within the meteorological domain

include inverse distance weighting (IDW), thin plate

splines and the kriging family, and should thus be within

the capability of the intelligent assistant.

This walkthrough is presented in Fig. 9. On start up,

the program presents the user with a desktop and the

informatics module and requests that the precipitation

data is loaded (Fig. 9a). After loading it appears in the

statistics frame pop up window and can be reviewed by

scrolling up or down. The first question asked is to

determine the user’s purpose, in this scenario the user

requires a surface that can be queried for analysis and the

[no] button is subsequently pressed. The following

question determines if we wish the surface to be smooth

or bumpy and the answer given is [no]. As Fig. 9

indicates, supporting figures appear within the infor-

matics module automatically as the user progresses from

question to question. The results of automated statistical

checks are also revealed to the user on an unprompted

basis in the statistics frame. The next question regards

whether the precipitation is related to another variable

(Fig. 9b). After further dialogue, including questions

relating to whether the covariates are locally or globally

related to the dependent variable, here elevation and

precipitation, the system concludes with a recommenda-

tion and the most and least appropriate of the available

methods are put forward. In this case universal kriging

was suggested as the method most appropriate (Fig. 9c).

This will be used under anisotropic conditions, when the

semi-variogram is a function of both the distance and

the direction of the distance intervals.

It is interesting to note that, in the case of this

walkthrough, the user was not a domain expert but was

rather trialling the passage of a naıve user through the

software. The user sensibly sought further help of the

software in determining mathematically whether there

ARTICLE IN PRESS

Class Javaawt.Window

ClassMyLine

ClassTestExpert

Class javax swingJComponent

Class Java lang.Object

ClassDataSetLoader

ClassSimpleIOClass Jess

Rete

ClassGeostatPanel

ClassNormalCurve

Class JessMain

ClassScattergramPanel

Class JessConsole

Class Javaawt.Frame

Class javaxswing.JPanel

Class Javaawt.Component

Class Javaawt.Container

ClassStatisticsModule

ClassConsolePanel

Java+Jess Package Class Hierarchy

Jess Class

Java Class

Key

ProjectClass

Class

Fig. 8. Java classes.


was an association between precipitation and elevation.

Investigation of the data was performed in the

exploratory analysis pull down menu. Pressing the

Pearson’s correlation button in the exploratory analysis

menu indicated that these variables were correlated and

thus the [yes] button was pressed by the user. The choice

or local or global association was however a relatively

arbitary one. By pressing [yes] to local variation, the user

ended with the suggestion to use universal kriging, a

good choice in relation to the existing literature, but

could equally have selected the potentially less desirable

recommendation to use residual ordinary kriging.

To assist the user in determining a theoretical semi-

variogram, i.e. to ascertain which function is necessary

to fit the data, the parameter analysis module is used,

which calculates the experimental (empirical) semi-

variogram. The start of this process is shown in Fig. 9d.

6. Verification, validation and evaluation

Significant numbers of articles on verification and

validation of knowledge-based systems first appeared in

the AI literature in the early 1980s. The methods used

are usually a reflection of that particular domain. The

approach for this research was to adopt the definitions

used by Gonzalez and Barr (2000, p. 412–413), where

‘‘verification is the process of ensuring that the intelligent

system conforms to specifications, and that its knowledge

base is consistent and complete within itself. Validation is

the process of ensuring that the output of the intelligent

system is equivalent to those of human experts when given

the same inputs’’. To suit our research needs in full, we

also advocated a sub-division of the verification process

into what we term logical and semantic verification.

Logical verification was used to ensure that the system

had no defects in the code. Thus, in order to determine

whether or not the statistical analysis methods, normal

curve, experimental variogram and descriptive statistics

were valid, the output was logically validated against

known samples worked from current systems such as

GS+TM from Gamma Design Software (2002). Where

time allowed the testing strategy in general attempted to

test the functionality and correctness of the programmed

code by running it to test for: (1) defect detection, and

(2) reliability estimation. In general, we populated a test

data file using randomly generated data so that the test

environment could be run automatically.

Semantic verification was used to determine if the

system had in fact incorporated consistent and complete

expert domain knowledge. This was implemented

though a review of the decision tree structures. In

particular, we checked the rule base for consistency,

correctness and redundancy, against the knowledge

ARTICLE IN PRESS

Fig. 9. Walk through of interpolation advisor in action for a precipitation data set from Peru. A Precipitation data set is loaded and

first question is asked of user, B Question is asked that user is uncertain about so exploratory analysis tool is envoked, C System gives a

recommendation and D Advisor assists with selection of variogram model through illustration of experimental semi-variogram.


acquired from the expert. In this case, this consisted of

viewing the decision tree for rules that have the same

premises but differing conclusions, which would suggest

inconsistency. For instance, kriging would normally be a

prime possibility for interpolations used for numerical

analysis. If it did not exist at a ‘classification leaf’ within

the expected side of the decision tree, then a review of

the process was instigated. Additionally, a graph of the

decision tree was presented to additional experts for

review. An independent expert verified the correctness of

the decision tree so that, for each possible case, the

likelihood would be that the right action(s) would be

executed by the system and the most probable inter-

polation method inferred. Furthermore, redundancy

checks were carried out in order to ensure that the

decision tree was not capable of repeating rules. By

definition, a binary tree eliminates redundant rules and

premises. Nonetheless, in some instances the same

question/rule needs to be applied to different nodes,

with potentially different outcomes. For example, on

nodes 6 and 7, the question is asked ‘‘do you have a

guiding variable’’; if the answer to this is [no] for node 6,

then the decision tree classifies a recommendation of 2d

splines. If on the other hand the answer was [no] for

node 7, another question is asked ‘‘do you know of any

significant trend in the data?’’ for which if the answer was

[no] and the data were of sufficient density a recom-

mendation of IDW would be presented. The decision

tree format easily allowed this kind of verification.

Running a series of validation scenarios on the system

and checking the output was an effective means of

proving that the interpolation method reasoned to be

appropriate by the system for a particular data set was

equivalent to that decided by a human expert. Previous

exhaustive empirical testing of methods for interpolating

daily temperature across Britain (Jarvis and Stuart,

2001a) provided a good base for comparisons bet-

ween the recommendations made by the system and


techniques actually known to be the most accurate for a

particular data set. In this case the system advised either

partial thin plate splines or de-trended (i.e. residual)

ordinary kriging on the basis of several days’ data

presented to it, which were indeed the two methods

performing best under comprehensive testing.

As Worboys (1995, p. 289) notes however, ‘‘Ulti-

mately, the success of any computer system rests on

whether it can be effectively used by people’’. No matter

how much thought has been given to the design of a

program to select an interpolation method and methods

of eliminating theoretical problems, it is only when

it is tested by the user that ‘‘real’’ problems appear.

For this reason, heuristic evaluation (Nielsen, 1994)

was adopted and used with extensive scenario walk-

throughs as a combined method for structuring a further

critique of the system as a whole. As part of this process,

the software was presented to six evaluators who

independently reviewed it for functionality or usability

problems. The evaluators ranged across a broad

spectrum, from the casual analyst to the expert in

interpolation selection to the interface designer. These

users employed the system, without help, to determine

the most appropriate interpolation method for a daily

rainfall data set from Peru. At the end of the

walkthrough, users were asked to complete a question-

naire formalized on relatively simple and general

heuristics. More specifically, it was composed of opinion

and attitude type questions divided into categories

examining content and cognition (5 questions), ease of

use (5 questions) and suitability for purpose (7

questions). A brief summary of the outcomes of the

evaluation exercise is outlined in Table 5. Within this

table, the values indicate overall measures of usability

based on total un-weighted scores from the question-

naires. The scores ranged from 1–7 where [1]=[Strongly

Disagree], [4]=[No Opinion], [7]= [Strongly Agree] and

[2,3,5,6] bounding measures.

In general, the users thought the system to be effective

in supporting them to select an interpolation method for

the presented meteorological data. Of note in the test,

were low marks for questions 3, 7, 10, and 14. Question

[3] asks if ‘‘the exploratory statistics module has

Table 5

Usability testing and evaluation

User type Casual

user

Expert

user

Interface

designer

Content and cognition

(5 questions)

26/35 22/35 21/35

Ease of use (5 questions) 28/35 16/35 20/35

Suitability (7 questions) 38/49 34/49 30/49

Scores are totals per group of questions, where each individual

question has a maximum score of 7.

sufficient analysis methods’’. The experts agreed that

the system could benefit with more ESDA and gave a

low score of 27; while the other users were content with

what was available ð57Þ: For questions [7] and [10] ‘‘I

believe I became productive quickly using this system’’

and ‘‘I am able to efficiently complete my work using

this system’’, again the experts gave a low score 27; while

the other users gave 57: Comments from the casual user,

such as ‘‘easy to use and navigateybased on windows,

which I’m used to’’ suggest in part that the problem with

interpolation experts is that they have set patterns of

decision making, access to familiar and tested tools, and

prefer to retain a wider sense of control over their choice

of interpolator. For question [14] ‘‘the system caters for

a generic user’’, again the scores are biased by the low

marks given by the experts 27: These interpolation experts

were again not entirely confident regarding the rainfall

patterns in Peru, but in their case they were aware that

further investigation was required on this matter and

that the software could have been more helpful in this

regard.

From the initial walkthrough that demonstrated the

visual panes of the software (Section 5) it is apparent

that automated diagnosis and ESDA are required where

the user is not confident of the spatial characteristics and

associations within the application domain. This reflects

the more negative comments received in the usability

and evaluation testing of a need to add more ESDA

functionality, and to further complete the help and

explanation modules to enhance the role of the system as

a tutor. Further ‘‘quirks’’ in the interface such as lack of

a start button and progress bars were highlighted. The

positive aspects were that the system ‘‘does enhance

understanding’’ and was generally accepted as being an

effective tool in supporting the user, getting an overall

score of 69%.

Overall, the results from the usability testing and the

process of scenario validation suggested that the

individual modules used in the prototype are comple-

mentary when used together, and encourage users to

become more aware of the properties of their data set.

This enrichment and understanding should improve the

quality of decisions in the task of selecting interpolation

methods that are most and least appropriate while at the

same time providing further functionality in supporting

the user to determine correct parameters for those

methods.

7. Concluding discussion

Based on a Graphical User Interface (GUI) designed

in Java, we developed a stand-alone ‘‘loosely coupled’’

prototype interpolation advisor which presents to the

user the most and least appropriate method for a given

task and data set. The prototype software that we


demonstrate is able to assist a variety of users in their

need to select an appropriate interpolation method,

without the comprehensive numerical testing of multiple

and perhaps inappropriate options. It also provides

further functionality in supporting the user to determine

correct parameters for those methods.

Having no dependency on other GIS program code to

run routines or statistical functions for the interpolation

selection process, the software is flexible in offering

assistance to a broader audience using a wide variety of

GIS packages for the subsequent interpolation itself.

The use of Java also allows for the ease of further

development of the system, owing to the rapidly

increasing number of publicly available classes for

geographical functions (e.g. Bastin et al., 2002; Faul-

kner, 2002). A process of logical and semantic verifica-

tion ensured the consistency of the knowledge encoding

within the prototype software. Logical verification is a

crucial part of the software cycle and is often overlooked

because of time and effort constraints (Arnow and

Weiss, 2000). Semantic validation was also achieved, by

comparing the methods suggested by the system with

methods actually chosen by the experts working with a

temperature data set.

A broader user evaluation process suggested that the

functionality implemented within the software was of

value to the general GIS user. Moreover, the users

appreciated the complementary nature of the individual

modules as implemented in combination. The system

clearly benefited from using multi modal channels to ‘re-

contextualize’ the knowledge held by the system. In

doing so it appeared to provide a richer structure of

references between logically related information, which

encouraged users to become more aware of the proper-

ties of their data set. This enrichment and understanding

should improve the quality of decisions in the task of

selecting interpolation methods that are most and least

appropriate. The functions implemented were also felt to

be useful to the general GIS user. Still, it is noted that

the specialist requested more techniques for ESDA such

as the enhancement of the parameter module and

graphical data summaries. Walkthroughs also high-

lighted the dangers of the software where the user was

not only new to interpolation, but was also not confident

regarding the application domain within which they

were working. Again, further ESDA or automated

analyses are suggested by this finding. Many of these

ESDA facilities are probably still best provided in

specialist software and linking the system to such

packages could be easily implemented, providing a

useful addition to the system functionality for the

specialist. In consideration, for the system to operate

across all platforms, the development of the prototype

as a stand-alone package is perhaps more worthwhile.

Further validation and evaluation exercises in different

user domains will be needed to test the system more

comprehensively once the recommended ESDA exten-

sions to the software are developed.

Outstanding research questions relate particularly

how to best employ the user and their understanding

within the decision making process, which reflect both

the context within which the software is being used and

the expertise of the users themselves. Use of the system

as an intelligent tutor, initially over the WWW but

perhaps better embedded within GIS such that the

difference between expected and learner pathways

through the decision making process can be highlighted

(e.g. Cheikes et al., 1998), suggests one profitable role

for multi-modal information and user driven explora-

tory analyses. Those who find the nature of interpola-

tion bewildering and the selection process problematic,

could be taught using the system. This develops the idea

of ‘‘rethinking’’ new ways to teach GIS (Fotheringham,

et al., 2000, p. 246). However, to comprehend confusing

context by visual experiences could be ‘‘influenced

(biased) by what we expect to see’’ (Mark, 1989, p.

551), thus, such visualizations must firstly be relevant to

the task, and secondly, for the user who is unaware of

interpolation terminology, must be of a type they expect

to see. This sounds contradictory, and suggests that

selecting such images in support of the interpolation

selection process is a difficult practice and an area

requiring further investigation. In contrast, the increas-

ing use of modular GIS components arranged for user-

specific tasks suggests the use of stronger and potentially

hidden structures to assist with the design of appropriate

workflows for particular tasks (Mineter et al., 2003). In

some contexts the software may need to fulfil completely

a function that normally requires human expertise, or it

may play the role of an assistant to a human decision

maker. The design of adaptable software (Grundy and

Hosking, 2002), with support that reduces as the user

gains familiarity with the system, is also an intriguing

prospect for the future.

Finally, a number of internal, expert systems and

knowledge structuring issues remain to be tackled.

Among these is to address the benefit of being able to

supply the user with a value that enumerates how

confident the system is of the appropriateness of the

method selected. This could be achieved in a future

system by summing the salience factors implied when a

rule fires (Giarratano and Riley, 1989), in a combination

with a move away from the strict decision tree structures

for knowledge used in this prototype. For the current

system, given the expert system implementation via Jess,

it would be a relatively straightforward enhancement;

however, this method only implies the importance of a

rule. Similarly, in its present prototype form the advisor

does not allow for stepping back and forth between

questions, nor does it allow for fuzzy answers. It asks of

the user to make a judgement that is either [yes] or [no],

and the introduction of fuzzy evidence is a further


direction for future work. In taking a broader view,

future work should focus on adding more meaning-

ful forms of quality assurance. Other future directions

should consider learning from experiences rather

than solely from theoretical knowledge. Case-

based reasoning matches new problems to ‘‘cases’’

adapting solutions from the past to current situations

(Giarratano and Riley, 1989). Further, cases are useful

for interpreting open-ended or ill-defined concepts and

thus fit well to task of selecting an interpolator.

Furthermore, it is possible to use a case-based system

as a way for a community of experts to pool their

knowledge.

Defining and encoding the process of how an expert

chooses an interpolation method is a complicated

undertaking. While there is a broad range of expert

knowledge on the subject of interpolation, there is a lack

of software to support the user in selecting an

appropriate method that balances expert knowledge on

context and procedures with data analyses. The proto-

type tool demonstrated here goes some way towards the

goal of developing a robust and comprehensive ‘in-

telligent’ advisor for selecting an interpolator, but its

development also raises many further research questions

for the future in the process.

References

Arnow, D., Weiss, G., 2000. Introduction to Programming

Using Java, An Object Oriented Approach. Addison-

Wesley, Longman, New York, NY 805pp.

Bastin, L., Fisher, P.F., Wood, J., 2002. Visualizing uncertainty

in multi-spectral remotely sensed imagery. Computers &

Geosciences 28 (3), 337–350.

Bonham-Carter, G.F., 1994. Geographic Information Systems

for Geoscientists: Modelling with GIS. Pergamon, Oxford

398pp.

Brezillon, P., 1999. Context in problem solving: a survey. The

Knowledge Engineering Review 14 (1), 1–34.

Bucher, F., 1998. Using extended exploratory data analysis for

the selection of an appropriate interpolation model. In:

Onsrud, H. (Ed.)., Geographic Information Research:

Trans-Atlantic Perspectives. Taylor and Francis, London,

pp. 391–403.

Cameron, M.A., Abel, D.J., 1996. A problem model for spatial

decision support systems, advances in GIS research II:

Proceedings of the Seventh International Symposium on

Spatial Data Handling, London, pp. 89–99.

Cheikes, B.A., et al., 1998. Embedded training for complex

information systems. Lecture Notes in Computer Science,

vol. 1452, pp. 36–45.

Collins, F., Bolstad, P., 1996. A comparison of spatial

interpolation techniques in temperature estimation. In:

NCGIA Third International Conference/Workshop on

Integrating GIS and Environmental Modeling, Santa Fe,

NM, USA.

Cooper W., 2001. Interpolation selection: supporting the user

with an intelligent system. M.Sc. Thesis, The University of

Edinburgh, Edinburgh, UK, 160pp.

Declercq, F.A., 1996. Interpolation methods for scaterred

sample data: accuracy, spatial patterns, processing time.

Cartography and Geographical Information Systems 23 (3),

128–144.

Dimitrakopoulos, R., 1993. Artificially intelligent geostatistics:

a framework accommodating qualitative knowledge-infor-

mation. Mathematical Geology 25, 261–279.

Duda, R.O., Shortliffe, E.H., 1983. Expert system research.

Science 220, 261–268.

Durkin, J., 1994. Expert Systems: Design and Development.

Prentice-Hall Inc, Englewood Cliffs, NJ 800pp.

Dykes, J.A., 1997. Exploring spatial data representation with

dynamic graphics. Computers & Geosciences 23 (4),

345–370.

Egenhofer, M.J., 1996. MultiModal spatial querying, advances

in GIS research II. Proceedings of the Seventh International

Symposium on Spatial Data Handling, London, pp.

785–799.

Erdmann, M., Studer, R., 1998. Use-cases and scenarios for

developing knowledge based systems. Proceedings of the

15th IFIP World Computer Congress, WCC’98, Conference

on Information Technologies and Knowledge Systems, pp.

259–272.

Faulkner, B.R., 2002. Java classes for nonprocedural variogram

modeling. Computers & Geosciences 28 (3), 387–397.

Forgy, C.L., 1982. Rete: a fast algorithm for the many pattern/

many object pattern match problem. Artificial Intelligence

19, 17–37.

Fotheringham, A.S., Brunsdon, C., Charlton, M., 2000.

Quantitative Geography: Perspectives on Spatial Data

Analysis. Sage Publications, London 270pp.

Friedman-Hill, E.J., 1997. Jess, The expert system shell for the

Java platform. Online Report SAND98-8206, Sandia

National Laboratories, Livermore, CA.

Gamma Design Software, 2002. GS+TM Version 5, Geostatis-

tics for the Environmental Sciences.

Giarratano, J., Riley, G., 1989. Expert Systems—Principles and

Programming. PWS-Kent Publishing Company, Boston,

MA 624pp.

Gonzalez, A., Barr, V., 2000. Validation and verification of

intelligent systems—What are they and how are they

different. Journal of Experimental and Theoretical Artificial

Intelligence 407–420.

Goovaerts, P., 1999. Performance comparison of geostatistical

algorithms for incorporating elevation into the mapping of

precipitation. Proceedings of the Fourth International

Conference on GeoComputation, Mary Washington Col-

lege Fredericksburg, Virginia, USA.

Grundy, J., Hosking, J., 2002. Developing adaptable user

interfaces for component-based systems. Interacting with

Computers 14, 175–194.

Hayes-Roth, F., Waterman, D.A., Lenat, D.B., 1983. Building

Expert Systems. Addison-Wesley, Reading, MA 444pp.

Heuvelink, G., Webster, R., 2001. Modelling soil variation:

past, present, and future. Geoderma 100 (3–4), 269–301.

Hutchinson, M.F., Gallant, G.C., 1999. Representation of

terrain. In: Longley, P.A., Goodchild, M.F., Maguire, D.,

Rhind, D.W. (Eds.)., Geographical Information Systems:


Principles and Technical Issues. Wiley, New York, NY, pp.

105–113.

Hutchinson, M.F., 1995. Interpolating mean rainfall using thin

plate splines. International Journal of Geographical In-

formation Systems 9, 385–404.

Jarvis, C.H., Stuart, N., 2001a. A comparison among strategies

for interpolating maximum and minimum daily air tem-

peratures Part I: the selection of ‘‘guiding’’ topographic and

land cover variables. Journal of Applied Meteorology 40

(6), 1060–1074.

Jarvis, C.H., Stuart, N., 2001b. A comparison among stra-

tegies for interpolating maximum and minimum daily

air temperatures. Part II: the interaction between number

of guiding variables and the type of interpolation

method. Journal of Applied Meteorology 40 (6),

1075–1084.

Jarvis, C.H., Stuart, N., Cooper, W., 2003. Infometric and

statistical diagnostics to provide artificially-intelligent sup-

port for spatial analysis: the example of interpolation.

International Journal of Geographical Information Science

17, 495–516.

Kitanidis, P.K., 1999. Introduction to Geostatistics: Applica-

tions in Hydrogeology. Cambridge University Press, Cam-

bridge, UK 271pp.

Lypher, S., Stelzner, M., 1991. Graphical knowledge-based

model editors. In: Tyler, S.W. (Ed.)., Intelligent User

Interfaces. ACM Press, New York, NY, pp. 403–420.

Mark, D.M., 1989. Cognitive image-schemata for geographic

information: relations to user views and GIS interfaces.

Proceedings, GIS/LIS, Orlando, FL, pp. 551–560.

Maslyn, R.M., 1987. Gridding advisor: an expert system for

gridding algorithms. Geobyte 2, 42–43.

Mineter, M.J., Jarvis, C.H., Dowers, S., 2003. From stand-

alone programs towards GRID-aware services and compo-

nents: a case study in agricultural modelling with inter-

polated climate data. Environmental Modelling and

Software 18, 379–391.

Mitas, L., Mitasova, H., 1999. Spatial interpolation. In:

Longley, P.A., Goodchild, M.F., Maguire, D., Rhind,

D.W. (Eds.)., Geographical Information Systems—Princi-

ples and Technical Issues. Wiley, Chichester, pp. 481–492.

Nielsen, J., 1994. Usability Engineering. Morgan Kaufmann

Publishers, CA, USA, 362pp.

Oliver, M.A., Khayrat, A.L., 2000. A geostatistical investiga-

tion of the spatial variation of radon in soil. Computers &

Geosciences 27 (8), 939–957.

Shannon, B., et al., 2000. JavaTM 2 Platform, Enterprise

Edition: Platform and Component Specifications. Addison-

Wesley Professional, Reading, MA 800pp.

Shneiderman, B., 1998. Designing the User Interface: Strategies

for Effective Human Computer Interaction. Addison-

Wesley, Reading, MA 640 pp.

Worboys, M.F., 1995. GIS: A Computing Perspective. Taylor

& Francis, London 392pp.

Documents

AJava-basedintelligentadvisorforselectingacontext- - …kclarke/Geography232/cooper_jar… · · 2005-02-02Computers&Geosciences30(2004)1003–1018 AJava-basedintelligentadvisorforselectingacontext-appropriatespatialinterpolationalgorithm$,$$